A short retrospective from a sprint where one coordinator agent supervised individual coding agents working on consecutive, dependency-linked issues. Generalized from a real run; concrete examples linked at the bottom.
Two LLM agent layers running against the same repo:
- Coordinator agent — owns sprint-level state. Reviews plans, reviews completion reports, files follow-up issues, edits the canonical sprint doc.
- Coding agents — one per issue. Read the issue body, sprint plan, ADRs, governing instructions (e.g. a repo-level
CLAUDE.md), and prior PR diffs. Propose a plan with acceptance criteria. Implement in an isolated branch/worktree. Produce a completion report aligned with the AC set.